Latent Brain
Small neural systems that learn state transitions, simulate limited action sequences, and evaluate their proposals against actual outcomes.
Research prototype · v0.4 · experimentalLearn, simulate, verify
Latent Brain is an open-source, CPU-first research project. It investigates whether small neural systems can learn from state transitions, retain information within an episode, predict short possible futures, and use those predictions to choose actions.
The published research covers two separate experiments: FlyBrain v0.3 navigates small mazes. The v0.4 reward lab trains a separate neural policy in a tiny synthetic debugging workflow. The training environments and model checkpoints are not the same.
Navigation in small mazes
FlyBrain combines recurrent memory (GRU), learned preferences for four actions, a prediction of the next hidden state, and a choice among 1, 2, or 4 simulated steps. Here, the choice of depth was trained on heuristic difficulty labels rather than evidence of optimal compute allocation.
Three training seeds were evaluated on 30 previously unseen 7×7 mazes each. A route counts as successful only if it reaches the goal during separate execution in the actual test environment ; the model’s prediction alone does not count.
| Method | Mazes solved |
|---|---|
| FlyBrain · adaptive depth | 77 / 90 |
| Same policy · fixed depth 1 | 74 / 90 |
| A* · learned transition model | 90 / 90 |
The aggregate advantage of three mazes was not consistent across the three seeds. A reliable benefit from adaptive thinking depth has therefore not been demonstrated. The A* comparator performed better in this toy benchmark.
Verified rewards in a synthetic test
A separate small GRU policy learns the actions inspect, test_a, test_b and report. Learning signals come from verified outcomes in the project’s toy test environment. An environment-owned ledger accepts reward receipts only once; the model cannot issue its own receipts.
Each seed used 250 training episodes. Evaluation covers three seeds with 80 unseen synthetic cases each, totaling 240 cases. An external interface may only propose actions; it cannot execute real changes.
| Method | Cases solved |
|---|---|
| Untrained neural policy (weak baseline) | 0 / 240 |
| Policy trained on verified toy rewards | 240 / 240 |
| Same policy · fixed depth 1 | 240 / 240 |
| Handwritten rule: inspect → test → report | 240 / 240 |
The neural policy learned this very small workflow. However, a simple non-neural rule achieves the same result. Neither better problem-solving ability nor a benefit from adaptive thinking depth has been shown. This is not a test involving real source code.
What the models do — and do not do
The published components explore recurrent state, learned action scores, and bounded predictions. Local inference can run on the CPU. An optional HTTP interface runs on the local machine and returns proposals, not proof of execution.
In v0.3 maze training, action examples come from verified demonstrations. The separate v0.4 reward model learns from rewards issued by the synthetic environment. These different learning methods must not be combined into a single performance claim.
What these results do not prove
The tests demonstrate behavior in narrowly constrained model worlds. They do not establish general reasoning or improvements to large language models. There were no real code repairs and no demonstrated integration with a production coding agent.
The maze and debugging experiments must be evaluated separately. Claims of practical benefit would require, among other things, new task families, stronger simple baselines, more independent seeds, and comparisons with equal compute budgets.
Transparency and evidence status
The numbers on this ARKEIN page come from published GitHub materials. We editorially checked them against the README, research documentation, and v0.4 evaluation report, but we have not independently rerun the experiments for ARKEIN. “Verified” refers to execution and checking inside the documented toy test environment, not an external audit or execution of real source code.
Sources reviewed: 18 October 2026 · GitHub branch main, commit 6907038.
Read and verify the sources yourself
The project is open-source under Apache 2.0. Start with the GitHub repository and its installation and reproduction instructions. The dedicated Latent Brain project website provides an additional overview.
README / Project status
FlyBrain v0.3 / Protocol
Verified Cookie Reward Learning v0.4
Evaluation report v0.4 / JSON
Earlier experimental methods (v0.1)